Tobacco farming is one of the main economic activities in agricultural areas of countries such as Brazil, China, and India [1]. During the last decades, a large number of technological innovations in this specific crop have been developed in diverse fields such as machinery, cultivation techniques, and drying techniques. Some advances in tobacco cultivation are the improvement of the health and quality of the tobacco plants by means of genetic modifications [2], the utilization of more effective fertilizers as a consequence of studies of their influence in tobacco growing and quality [3], or the employment of new machinery and techniques to irrigate and collect tobacco plants [4]. Furthermore, the tobacco drying process has been studied and improved by employing new drying strategies and new machinery in order to increase the quality of the dry tobacco and the overall energy efficiency of the process [5].

The tobacco drying process requires exhaustive monitoring of its representative variables to optimize it [6]. Monitoring systems are usually employed to acquire and store the most useful variables of a process. In the literature, these systems have been proposed to monitor atmospheric phenomena such as air pollution [7], livestock farms such as aviary production systems [8], or environments including rivers [9], oceans [10], and ecological systems [11]. In the field of agriculture, monitoring systems have been proposed for indoor locations such as greenhouses [12–14], and for outdoor agricultural processes including wine harvesting [15], olive growing [16], or any generic process [17].

Data stored by monitoring systems have been used to analyse the connection among the system's variables and to obtain a model of the monitored process [18]. These models can be defined by means of fuzzy-logic, probabilistic estimators, or Artificial Neural Networks (ANN) among others. Using fuzzy-logic, Schulz et al. [19] modelled water flow, Barros et al. [20] analysed different demographic and environmental models, Malins et al. [21] modelled the spatial extent of salinity on a farming land, and Papantoniou et al. [22] proposed a model of wind power for farm applications. Using probabilistic estimators, Cornford et al. [23] and Ng et al. [24] proposed different models to forecast precipitations, Holland et al. [25] modelled animal populations, and Trombe et al. [26] modelled offshore wind power fluctuations. Using ANN, Recknagel et al. [27] modelled algal growing, Smith et al. [28] predicted the air temperature, and Singh et al. [29] modelled water quality.

The aforementioned processing data tools have also been employed to control and predict variables in agricultural drying processes. Using fuzzy-logic, Yliniemi et al. [30–32] designed the control logic for a rotary dryer and Li et al. [33–35] developed and tested a control algorithm for microwave drying of vegetables. Employing probabilistic estimators, Banga et al. [36] designed a stochastic control algorithm using temperature and relative humidity as input variables. Using ANN, Yliniemi et al. [30,31] designed a system to identify the dynamics of the drying process in a rotary dryer, and Movagharnejad et al. [37] proposed a tomato drying model which fitted the experimental data more accurately than other mathematical models.

A system to estimate and predict the temperature and relative humidity in the tobacco drying process is proposed and tested in this work. This system acquires temperature and relative humidity data in different points of a tobacco dryer, and processes this data with an ANN in order to estimate and predict their value in other locations or in the future. The methodology of the article was composed by four stages: (i) a wireless sensor network (WSN)-based monitoring system was deployed in the drying installation to capture input variables, (ii) an artificial neural network was developed to estimate output variables using input variables, (iii) different experiments were performed to analyse the proposed methods, and (iv) experimental results were analysed to evaluate the proposed method.

In light of the results, it is easy to realize that the use of the ANN results in a more efficient estimation and prediction task performance. For instance, for temperature estimation when there are no direct temperature readings in the system, the neural network offers a sufficient performance in the surroundings of 6–7% mean error, which is more than enough for an approximation; but even more, when there is temperature data, the ANN is capable of being between two and three times more precise than the interpolation method considered. When estimating the humidity, the results are even more meaningful, because even if the absolute mean error registered is around 5%, this error is between five and six times lower than the one obtained by interpolation.

The prediction task is slightly more complex, since, depending on the conditions, the performance of the ANN could greatly vary when compared to interpolation. Due to magnitudes varying slowly, it has to be considered that for small prediction horizons it might be better to directly interpolate the values. However, it has been shown that as the horizon increases, the neural network is capable of a more accurate prediction.

Estimations provided by the proposed system can be employed to detect sensors' failures and to recover their measurements when failures happen. Sensor failures can be detected comparing a threshold with the absolute difference between the estimated measurement obtained by the estimation method and the sensor measurement. If the absolute difference is higher than the threshold, then a failure has occurred. In this case, and until the sensor problem is solved, the measurements of this sensor can be replaced by the estimation provided by the ANN using the rest of the sensors measurements as input data. Moreover, this data recovering application can be used to recover data when sensors have other type of failures, such as sensor signal corruption or sensor signal loss, which are two typical problems in WSN [40,41].

Another application of the proposed estimation methods is the design optimization of data acquisition systems. To this end, a set of sensors must be deployed in several significant points of the monitored system. After that, the proposed estimation methods can be employed to discard the sensors whose measurements can be more accurately estimated using the rest of the deployed sensors with a low error. One advantage of these designed systems is their economy due to discarding the sensors whose measurements can be estimated with the other sensors of the system. Another advantage is that the measurements of sensors sited in a point with dangerous or hostile conditions that can quickly degrade the hardware can be replaced with the estimated measurements calculated from data of other sensors sited in safer places with better conditions.

Proposed estimation and prediction methods can be used to model processes. Models establish relationships among the variables of a process and predict the future evolution of these variables. They have been applied in the literature to plant growing processes [42–46], vegetal evapotranspiration processes [19,47,48], or drying processes [37,49,50] among others. One way to establish the relationships and to predict the evolution of processes' variables is using ANNs, which have been used to model algal growing [46] or drying processes as in the carrot drying process [51] and in the tomato drying process [37]. Process models are usually developed as a part of model-based predictive control (MPC) algorithms. MPC are multivariable control algorithms which use prediction methods to estimate the future state of the system and an optimization cost function in order to calculate the optimum control signals taking into account the predicted future state [52–63]. ANN-based prediction methods as that proposed in this article can be used in a MPC algorithm as the prediction method. ANN-based MPC algorithms have been proposed to control industrial applications [59–61], aircraft navigation [62], and laboratory processes [63].

Another application of the proposed data prediction methods is the improvement of alarm systems' performance. A typical problem in alarm systems is the high rate of false positives and false negatives. One cause of this problem is that alarm systems only take into account present data, or past and present data, so the incorporation of predicted future information about the system can improve the precision of the alarm system [64].

Finally, the proposed system's predictions can be easily employed to anticipate decisions about the drying process. The operator can see a forecast of the variables of the tobacco dryer in the future on the basis of the present conditions and also under a series of alternative conditions, making the process of taking the most appropriate decision to reach the desired values in the future much easier.

Tobacco drying has been traditionally a manually controlled activity, relying solely on experts' craftsmanship, and has been considered almost an art; however, as in many other areas of agriculture, new technologies have appeared as an easy solution to greatly improve the results and performance of the production processes while reducing costs at the same time. Specifically, introduction of sensor networks and intelligent automation systems represents a strong trend, as shown in the state of the art of drying processes. The performance of these systems depends heavily on the accuracy of the data provided by the sensor network, and as such, the better the data provided, the better the results.

The work presented in this paper has shown that by using a fitting ANN it is possible to accurately estimate the values of the different environmental magnitudes employed to control the drying process at different spatial points. The numerical results obtained present very low errors, which at less than 2% could even be below some sensors' accuracy.

Additionally, the system proposed is also capable of predicting the values of those environmental magnitudes with different time horizons. In this case, the results are dependent on the specific time horizon considered. For nearby time horizons, the mean errors given by interpolation are similar to those of the ANN system; however, for distant time horizons, the mean errors given by the ANN are much lower than those obtained by interpolation.

Finally, it is worth mentioning that both capabilities of the presented system are potentially very useful for improving the performance of automatic control of tobacco drying processes while at the same time reducing costs, by increasing the quality and quantity of available data.